Triage tickets and save drafts

This guide is for a support team that wants to classify a ticket and save a reply draft in its ticketing website. GLiNER2.5 Decide suggests the team and priority. Ornith uses Computer Use to operate the selected macOS window. A person checks the ticket before sending the reply.

To route tickets by using GLiNER and Cua Driver without Ornith, follow Route tickets with GLiNER and Cua Driver.

Start with one ticket that your team authorizes for a draft-only run. You can use the workflow without adding model code or an API to the ticketing app. A separate controller connects the steps to an inbox.

Prepare your Mac and ticket

Use an Apple Silicon Mac with macOS 15 or later. Open the ticketing website in one window and select the ticket that you will process. Record its ticket ID, status, customer message, and any order or account context needed for triage. List the actual team, priority, and tag choices in your app. Decide which outcomes require human review before an app action.

Ornith's 4-bit model needs at least 32 GB of unified memory; 48 GB is the conservative recommendation. Check that your mere.run command supports classification:

mere.run text classify --help

For source build prerequisites, see Getting started.

If classification is unavailable, build mere.run from source and use that binary in the terminal where you run the plugin:

git clone https://github.com/sawfwair/mere-run.git
cd mere-run
swift build -c release
export PATH="$PWD/.build/release:$PATH"
mere.run text classify --help

To install the models and Pi, the local agent runner, run:

mere.run model capabilities
mere.run model pull text-classify-gliner25-decide
mere.run model pull text-agent-ornith-35b-mlx-4bit
mere.run agent onboard --install-pi

To install Computer Use and check the selected-window setup, run:

mere.run plugin install mere-computer-use --yes
cua-driver permissions grant
mere-computer-use doctor --model text-agent-ornith-35b-mlx-4bit
mere-computer-use windows

Grant CuaDriver.app Accessibility and Screen Recording permissions in macOS System Settings if doctor reports that they are missing. The plugin can start the local Ornith API for a run. It does not install the models. For driver setup and supported models, see the Computer Use plugin guide. If doctor reports that the installed CLI cannot process Ornith screenshots, use the source build described in this guide.

Classify the selected ticket

Create a private work directory outside a source repository:

umask 077
mkdir -p "$HOME/mere-run-support"
chmod 700 "$HOME/mere-run-support"
cd "$HOME/mere-run-support"

Create the ticket-classification.json file in that directory. Use the ticket's actual message and relevant context as text. Replace each uppercase label with a team or priority that your app accepts. Include every team and priority that the controller may select. No action tells your controller to skip the ticket; it does not need to be a team in the app. The following file shows the request shape. It contains placeholders, so edit it before running the command:

{
  "text": "REPLACE_WITH_TICKET_MESSAGE_AND_RELEVANT_CONTEXT",
  "tasks": [
    {
      "name": "team",
      "labels": ["TEAM_LABEL_A", "TEAM_LABEL_B", "No action"]
    },
    {
      "name": "priority",
      "labels": ["PRIORITY_LABEL_A", "PRIORITY_LABEL_B"]
    }
  ]
}

To check that the request fits the model and then get suggestions, run:

mere.run text classify --input ticket-classification.json --preflight --pretty
mere.run text classify --input ticket-classification.json --pretty

Review both selected labels and scores. The scores are not calibrated probabilities. Skip a resolved ticket or a No action result. If the message does not support the suggested team or priority, leave the app unchanged and send the ticket for human review. For task limits and more request options, see GLiNER2.5 Decide.

Save a draft in the selected window

In the mere-computer-use windows output, find the process ID and window ID for the window that contains your selected ticket. Use the team and priority that you accepted after reviewing GLiNER's output. In the following command, replace WINDOW_PID, WINDOW_ID, TEAM_LABEL, and PRIORITY_LABEL with those values:

SUPPORT_TASK="Work only on the selected ticket.
Suggested team: TEAM_LABEL; priority: PRIORITY_LABEL.
Recheck the ticket ID and message before changing anything.
Set those fields only if the visible request supports them.
Add relevant tags. Write a reply and save it as a draft.
Do not send a reply or change ticket status. Observe the final state."
mere-computer-use plan \
  --pid WINDOW_PID \
  --window-id WINDOW_ID \
  --model text-agent-ornith-35b-mlx-4bit \
  --max-actions 12 \
  --task "$SUPPORT_TASK" \
  --output ./support-ticket-run

The plan command creates ./support-ticket-run/run.json without changing the app. Review its window and task. To let Ornith operate that window, run:

mere-computer-use run ./support-ticket-run/run.json

The task text instructs Ornith to save a draft, but the prompt is not a technical block on the app's Send control. Use an account or app setting that prevents sending during the first run when your ticketing system provides one. Keep an operator present.

After the run, open the same ticket in the app. Check the team, priority, tags, draft text, unchanged status, and sent-reply count. Use a read-only API or audit log for an independent check when you have access. A model report or action count alone does not verify the final ticket state.

If the run stops, inspect the ticket before doing more work. The mere-computer-use resume ./support-ticket-run/run.json command displays the saved record; it does not replay the actions. If work remains, create a new plan only after checking whether a draft was already saved.

Process more than one ticket

The commands in this guide require you to review GLiNER's result and pass the accepted labels to Computer Use. The plugin does not discover or classify an entire inbox by itself. To connect the steps for your website, follow Adapt the support workflow. That guide covers ticket discovery, changes between observation and action, verification, resume, and a prompt for a coding agent.

Measure classification errors, completed actions, and total wall time on a set of tickets from your own workflow before increasing the scope.